Forecasting urban traffic requires consideration of both general temporal patterns and local dynamics, which depend on specific sensors. In this paper, we develop and evaluate a cluster-based forecasting architecture for the PEMS04 and PEMS08 traffic datasets. First, sensors are clustered using the DBSCAN algorithm based on robust descriptors (mean, standard deviation, and upper quantile statistics regarding traffic intensity, lane occupancy, and speed). Second, forecasting models are trained and compared across three operational horizons: 15, 30, and 60 min (H = 3/6/12 with 5 min granularity). The comparison includes Historical Average (HA), Global LSTM, CNN-LSTM, and Cluster-Aware LSTM. The results show that Cluster-Aware LSTM provides the best MAE on short and medium horizons (H = 3, H = 6), while HA remains the strongest on the long horizon (H = 12), indicating strong seasonal stability that extends beyond a single hour. Cluster-level confidence matrices reveal a systematic concentration of difficulties in groups where congestion prevails. Explainability diagnostics (accuracy and stability) demonstrate very stable attributions and stronger accurate signals in the cases of the most challenging clusters. Overall, the study supports a hybrid operational strategy: deep, cluster-adaptive forecasting for short- to medium-term horizons and a robust seasonal buffer for forecasting longer-term horizons.
Savchuk et al. (Tue,) studied this question.